Download www.ijecs.in International Journal Of Engineering And Computer Science ISSN:2319-7242

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Transcript
www.ijecs.in
International Journal Of Engineering And Computer Science ISSN:2319-7242
Volume 2 Issue 11 November, 2013 Page No. 3097-3100
Identification of level of resemblance between web
based documents
Surbhi Kakar1
Assistant Professor, GGSIPU
Abstract—One of the biggest challenges today on web is to deal with the “Big data” problem. Finding documents which are near duplicates of
each other is another challenge which is in turn brought up by Big data. In this paper the author focuses on finding out the near duplicate
documents using a technique called shingling. This paper also presents the different types of shingling that can be used. Further, a measure
called the Jaccard coefficient is discussed which can be used to judge the degree of similarity between the documents.
Index Terms—Big data, shingling, Jaccard Coefficent
INTRODUCTION
“Big Data” as the name suggests refers to a large amount of
data. Every industry today is enjoying the benefits as well as
dealing with the challenges brought up by the ever increasing
data of their own companies. Data mining is a process where
in it is possible to mine huge amounts of data in order to
expose hidden patterns. For instance, with the help of data
mining techniques today, it is possible for a business
organization to get insights into what their customer wants or
how he will react in future. Therefore, mining this Big Data
can help discover user’s hidden behavioral patterns and their
intentions too. [1] The major challenges with “big data” are
the increasing size, heterogeneity and velocity of data being
added to the space every moment [2].
The increasing size of data is resulting in storage overheads. A
serious issue is what to do with this data. What data should to
be discarded and what should be stored. The heterogeneity of
data on the other hand refers to the data coming from several
sources which might be unstructured or even incomplete. For
instance, data can be in various formats ranging from simple
records to geo spatial data.
The velocity of data refers to the rate at which data is
continuously being added on the web. Such data is becoming
difficult to analyze and interpret which is also taking a large
processing time.
Duplicate or near duplicate documents are the documents
which might be exact replicas of each other or might be
similar to one another. Finding out the near duplicates of web
documents in a scenario like today where the web is facing the
challenges of Big data becomes a serious issue to handle. This
is mainly because the users today are just not interested in
responses that are mere duplicates of each other. Also,
indexing such documents affects the storage and processing
time of the search engines. The resemblance between
documents can vary between 0 to 1. A “1” indicates that the
two documents compared for similarity are almost the same.
Whereas a “0” indicates a higher level of dissimilarity
between the two documents [4]
Shingling is a technique which can be used to find out the near
duplicate documents. It is based on creating contiguous
subsequences of a document of length q. These
shingles can be created by either using subsequences of
characters or words.
Jaccard Coefficient on the other hand, is a measure which
determines the similarity between two documents that are
represented as sets. This measure ranges from 0 to 1.
This paper defines the problem definition in the first section
which is to judge the level of resemblance between two
documents.
The second section describes about shingling as a technique
used to solve the problem. The section thereby discusses about
the Jaccard coefficient measure used to judge the similarity
between documents.
1. PROBLEM DEFINITION
One of the problems associated with “Big Data” is finding
duplicate or near duplicate documents as discussed previously.
As per [3] it has been estimated that as many as 40% of web
pages are duplicates of each other. The duplicity here refers to
documents which are textually similar and not semantically
Surbhi Kakar1IJECS Volume 2 Issue 11 November, 2013 Page No.3097-3100
Page 3097
similar. Some of the common examples of duplicate/near
duplicate web documents are mirror sites, news articles from a
common source, plagiarized documents. Mirror sites are sites
which are duplicated at a number of hosts to facilitate load
sharing. Plagiarized documents are the ones that may not be
exact copies of one another but it is a possibility that they
share large amounts of text which may not be in the same
order.
The documents which are exact replicas of each other are still
easier to find as compared to documents that are similar.
Finding such documents require the use of hash functions [8].
The items which are duplicates will have the same hash value
as compared to the dissimilar items. Therefore detecting near
duplicate pages is very important because it affects the search
engines time and space complexity when it has to index and
store duplicate web documents.
2. SHINGLING: A SOLUTION
Shingling as discussed above is a technique of creating
consecutive subsequences of tokens of a document. The
tokens here can refer to characters, words or lines. A
document can be represented as a set of string of characters.
Given a document d, we can create ‘t’ tokens of d which are
contiguous sequences of the text of d. Creating a q-shingle/qgram means creating a set of substrings of d which are of
length q[4][5]
Shingling can be done by creating tokens of characters as well
as words.
A. Shingling by characters
Let S(d1) be a set of document d1 which is defined as
“acdacef”. Then q-shingles based on letters for the above
document where q=2 will be {ac, cd, da, ce, ef}.
The elements appearing multiple times in the document are
only considered once while creating a shingle set.
around 9 whereas for documents like comparing emails for
similarity, the size q can be picked as 5.
3. JACCARD COEFFICIENT:
A MEASURE FOR RESEMBLANCE BETWEEN
DOCUMENTS
Jaccard coefficient is used to measure resemblance between
two documents. For two documents represented as sets A and
B, the Jaccard coefficient is the ratio of intersection of two
sets to the union of them[5][7]. It can be calculated as:
JC(A,B)=|A ∩ B| / |A U B|
Where JC(A,B) denotes the Jaccard Coefficient between two
sets A and B. Jaccard coefficients are therefore in a range of
[0,1].
For instance if A={3,4,5,6} and B={5,6,7,8} then:
JC(A,B)=2/4
This measure is used to judge the similarity between only
“textually” similar documents and not “semantically” similar
document. Two documents with a high Jaccard coefficient
denote a higher degree of similarity between them. But at the
same time keeping the size of shingle too small may bring
about a higher Jaccard coefficient which may not imply that
the two documents are similar.
5. IMPLEMENTATION
compare(n)
{
for i=0 to n
for j=0 to n
Set S1=getShingle(readFile(d1, q)
B. Shingling by words
Shingling by words is done by using words as substrings of a
document. If S(d1) contains a string:
“I did my work today”
Set S2=getShingle(readFile(d2,q)
findJaccardSimilarity(s1,s2)
}
//finds Jaccard Similarity between two sets
Then q-shingles based on words where q=3 will be:
“ I did my”, “did my work” “my work today”
In such a case, space used to store these shingles would be
O(q*w), where q is the length of the shingle and w are the
number of words in the document
C. Size of Shingles
While creating the shingle sets, the point to be focused upon is
what should be the size of q so that the probability for the
similarity between the similar documents is maximized as
compared to dissimilar documents. If q is kept too small, there
is a probability that most of the text in one document also
appears in the other document in spite of them being
dissimilar textually. As per [6], the size of q should be picked
large enough so that the probability of one shingle set
appearing in the other document is low. For comparing large
documents like research papers the size of q can be taken
findJaccardSimilarity(S1,S2)
return (S1 intersection S2) /(S1 union S2)
Here n is the no of documents to be compared for similarity.
getShingle method first reads the document d1 and d2 amongst
the list of documents being iterated and then makes shingles
by characters and words of length q. It does it by finding
consecutive substrings of length q for each document. It then
finds the Jaccard similarity between each set being iterated.
While making the shingle sets, the blank spaces can be split up
to form one blank space.
This algorithm makes O(n2) comparisons to compare each
document against the other. For instance, if there are 10
documents, each document will be compared with all the other
documents resulting in overall 45 comparisons. The space
Surbhi Kakar1IJECS Volume 2 Issue 11 November, 2013 Page No.3097-3100
Page 3098
complexity for the same would be O(q*w) as discussed
previously.
6. EXPERIMENT
Shingling as a technique was used by the author on a dataset
of 10 text files. The length of shingle was taken as 3, 4 and 5.
A java program was written to implement the same. Out of the
10 text files the Jaccard coefficients calculated for the first
four files are summarized in Table1, 2 and 3.
TABLE1
q=3
Similarity between
Similarity between
Similarity between
Similarity between
Similarity between
Similarity between
i: 2 j: 1
i: 3 j: 1
i: 3 j: 2
i: 4 j: 1
i: 4 j: 2
i: 4 j: 3
0.11809045226130653
0.023178807947019868
0.03070175438596491
0.01559792027729636
0.15834767641996558
0.022974607013301087
Fig. 1. Jaccard coefficients for q=3, 4 and 5
On experimenting the dataset with the algorithm, it was found
that on varying the values of q the Jaccard coefficient also
varied. For smaller values of q, Jaccard coefficient came out to
be larger which indicated a higher degree of similarity
between the documents. Whereas as the value of q was
increased the Jaccard coefficient decreased.
7. CONCLUSIONS
The paper discussed about the problem of “Big Data”
emerging at a speedy rate. It also focused on problems like
near duplicate detection of documents for which it described a
technique called shingling. In the later sections the author
brought about two approaches by which shingling can be
done.
As per the experiment conducted by the author, it was found
that for smaller values of the length of the shingle (q) the
jaccard coefficient between the documents was larger as
compared to when the value of q was increased.
q α 1/JC(A,B)
TABLE 2
q=4
Similarity between
Similarity between
Similarity between
Similarity between
Similarity between
Similarity between
i: 2 j: 1
i: 3 j: 1
i: 3 j: 2
i: 4 j: 1
i: 4 j: 2
i: 4 j: 3
0.09535452322738386
0.01461038961038961
0.018571428571428572
0.003372681281618887
0.13079470198675497
0.009389671361502348
TABLE 3
q=5
Similarity between
Similarity between
Similarity between
Similarity between
Similarity between
i: 2 j: 1
i: 3 j: 1
i: 3 j: 2
i: 4 j: 1
i: 4 j: 2
Similarity between i: 4 j:
3
0.07971014492753623
0.00967741935483871
0.011315417256011316
0.0016806722689075631
0.11201298701298701
0.003484320557491289
Note: Here i and j refer to the documents that are being
compared for similarity.
The statistical analysis of the above result can be stated as
below:
This implied that a higher Jaccard coefficient does not
necessarily indicate that the documents are similar. It can vary
depending upon what value of q is chosen. Therefore, the
value of q should be appropriately chosen in order to have an
accurate estimate of degree of similarity between the
documents.
8. LIMITATIONS
For very large values of n where n is the number of documents
to be compared for similarity, shingling alone as a technique
used can result in large processing and storage overheads since
it takes O(n2) comparisons to compare each document. As a
solution shingling can be clubbed with Minhashing or Locality
sensitive hashing(LSH) to improve the results. Locality
sensitive hashing does not compare all the documents for
similarity. It only focuses on the candidate pairs that are likely
to be similar for comparison. This reduces the comparisons to
be done while checking documents for similarity. Also LSH is
a better technique while searching documents in high
dimensional spaces.
9. FUTURE SCOPE
The author will extend the work discussed in this paper and
implement Minhashing and Locality sensitive hashing
techniques to improve the results in order for efficient near
duplicate detection of web pages.
REFERENCES
[1] Michael, Katina, and Keith W. Miller. "Big Data: New
Opportunities and New Challenges." Editorial: IEEE
Computer 46.6, pp.23, 2013
[2] Einav, Liran, and Jonathan D. Levin. “The Data
Revolution and Economic Analysis”.No. w19035. National
Bureau of Economic Research, pp.3-5, 2013.
Surbhi Kakar1IJECS Volume 2 Issue 11 November, 2013 Page No.3097-3100
Page 3099
[3] Manning D. Christopher , Raghavan Prabhakar , Schütze
Hinrich , “Introduction to Information Retrieval”, Cambridge
University Press, April 2009
[4] Broder, Andrei Z. "Identifying and filtering near-duplicate
documents." In Combinatorial Pattern Matching, Springer
Berlin Heidelberg, pp. 2-6, 2000.
[5] Broder, Andrei Z. "On the resemblance and containment of
documents." In Proceedings of Compression and Complexity
of Sequences, IEEE, pp. 3-6, 1997.
[6] Rajaraman, Anand, and Jeffrey David Ullman, “Mining of
massive datasets” Cambridge University Press, pp. 6, 2012.
[7] Andoni, Alexandr, and Piotr Indyk. "Near-optimal hashing
algorithms for approximate nearest neighbor in high
dimensions." In Foundations of Computer Science, 2006.
FOCS'06. 47th Annual IEEE Symposium on. IEEE, pp.120,
2006.
[8] Ye, Shaozhi, Ji-Rong Wen, and Wei-Ying Ma. "A
systematic study on parameter correlations in large-scale
duplicate document detection." Knowledge and Information
Systems 14.2, pp. 2-6, 2008
Surbhi Kakar1IJECS Volume 2 Issue 11 November, 2013 Page No.3097-3100
Page 3100
Document related concepts

Nonlinear dimensionality reduction wikipedia, lookup